A multi-factor influenced agricultural equipment inventory demand forecasting method

Abstract

The application discloses a kind of multi-factor influence agricultural equipment inventory demand prediction methods, steps are as follows: one, obtains inventory demand and influence factor historical data, influence factor historical data is preprocessed;Two, using empirical mode decomposition algorithm will inventory demand historical data be adaptively decomposed into multiple different scales intrinsic mode component imf And a residual component res;Three, using Fourier transform selects the imf with periodicity characteristics, res and corresponding imf are input into Prophet algorithm and are trained to model, according to this prediction model, the corresponding imf is predicted;Four, using principal component analysis selects the key influence factor of remaining imf, and the key influence factor is used as independent variable, input DRL-LSSVM and is trained to model, according to this prediction model, the imf is predicted;Five, the above two prediction results are added to obtain the final prediction result.The method effectively improves the accuracy of solving seasonal demand under agricultural equipment inventory demand prediction, improves the explainability of model.

Description

Technical Field

[0001] This invention belongs to the field of demand forecasting technology, specifically a method for forecasting the inventory demand of agricultural equipment influenced by multiple factors. Background Technology

[0002] With technological advancements and the modernization of production methods, it has become increasingly easier to mass-produce standardized agricultural equipment. As a result, agricultural equipment manufacturers are finding it increasingly difficult to gain a competitive advantage from products that are relatively abundant and lack differentiation. They are beginning to pay more attention to the importance of inventory cost control in warehousing while focusing on overall efficiency, and are placing higher demands on the accuracy and scientific nature of inventory demand forecasting.

[0003] Due to the seasonality of demand for agricultural equipment products, the difference in demand between peak and off-peak months is significant. To cope with peak production, it is necessary to establish "reserve inventory" in advance for some specialized components with long manufacturing cycles and complex structures. This inventory differs from order-driven inventory in ordinary supply chain management. Although it is related to the current order volume, it depends more on the influence of multiple factors or the outcome of game theory. Traditional forecasting methods have certain limitations in practical applications and are difficult to achieve satisfactory forecasting results in the face of highly unstable economic environments. Artificial neural network models, lacking corresponding economic theoretical support, cannot explain the economic meaning of the models. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a multi-factor-driven method for predicting agricultural equipment inventory demand. This method effectively improves the accuracy of inventory demand forecasting under seasonal demand conditions and enhances the interpretability of the model. Building upon existing forecasting techniques, it combines the Prophet algorithm and the DRL-LSSVM forecasting model to further improve the algorithm's efficiency and stability, providing new insights for future research.

[0005] This invention is achieved through the following technical solution:

[0006] A method for forecasting agricultural equipment inventory demand influenced by multiple factors includes the following steps:

[0007] Step 1: Obtain historical data on inventory demand and influencing factors from the departmental database, and preprocess the historical data on influencing factors.

[0008] Step 2: Using the empirical mode decomposition algorithm, the historical data of inventory demand is adaptively decomposed into multiple intrinsic mode components (IMF) of different scales and a residual component (RES).

[0009] Step 3: Use Fourier transform to select the imf with periodic characteristics, input res and the imf with periodic characteristics into the Prophet algorithm for training and modeling, and make predictions on the imf with periodic characteristics based on this prediction model.

[0010] Step 4: Use principal component analysis to select the key influencing factors of the remaining non-periodic IMFs, use the key influencing factors as independent variables, input them into DRL-LSSVM for training and modeling, and use this prediction model to predict the non-periodic IMFs.

[0011] Step 5: Add the prediction results from Step 3 and Step 4 together to obtain the final prediction result.

[0012] Furthermore, the influencing factors in step 1 specifically include: inventory level, order volume, raw material price, qualification rate, productivity, agricultural machinery subsidy policy, farmers' income, fuel price, arable land policy, and grain price.

[0013] Furthermore, the data preprocessing of the historical data of the influencing factors in step 1 specifically involves: normalizing the historical data of the influencing factors.

[0014] Furthermore, in step 2, the historical inventory demand data x(t) is decomposed using EMD, resulting in:

[0015]

[0016] In the formula, x(t) is the value of inventory demand at time t; imf is the intrinsic modal component, arranged in descending order of frequency; res is the residual component; and j is the number of intrinsic modal components.

[0017] Furthermore, in step 3, Fourier transform is used to select the imf with periodic characteristics. Specifically, Fourier transform is used to transform each imf to the frequency domain. The signal whose frequency domain graph is discrete is the periodic signal.

[0018] Furthermore, in step 3, the Prophet algorithm fits the trend, seasonal, and holiday terms in the time series and sums them to obtain the final prediction result. The Prophet algorithm is then trained and modeled, and the specific steps are as follows:

[0019] (1) Input the timestamps and corresponding values ​​of the time series and train the model;

[0020] (2) Input the length of the time series to be predicted into the trained prediction model;

[0021] (3) Output the future time series trend, including the fitted curve, upper bound and lower bound.

[0022] Furthermore, in step 4, principal component analysis is used to select the key influencing factors of the remaining non-periodic IMF. Specifically, principal component analysis is used to reduce the number of influences, remove unnecessary features, generate features with stronger explanatory power as much as possible without losing most of the information, and generate several principal components that retain most of the information of the original features, i.e., key influencing factors.

[0023] Furthermore, the specific steps of regression prediction based on DRL-LSSVM in step 4 are as follows:

[0024] (1) In step 4, an m×q principal component score matrix is ​​obtained according to the principal component analysis method. This matrix is ​​used as the input data, and the remaining imf, which does not have periodic characteristics, is used as the output data and input into the LSSVM model.

[0025] (2) The agent in the reinforcement learning algorithm selects a set of hyperparameters λ with probability P as the LSSVM algorithm, trains the algorithm on the training dataset, uses the prediction error as the reward value, and uses the policy gradient algorithm to update the policy. Finally, a set of hyperparameters with the smallest prediction error is selected to obtain the prediction model.

[0026] (3) Input the data to be predicted into the prediction model and output the prediction results.

[0027] Furthermore, the specific steps in step 5, which involve summing the prediction results, are as follows:

[0028]

[0029] In the formula, x(t') is the predicted inventory demand at time t'; imf j (t') is the predicted value of the j-th intrinsic mode component at time t'; res(t') is the predicted value of the residual component at time t'.

[0030] The beneficial effects of this invention are as follows:

[0031] (1) In view of the complexity of the problem of agricultural equipment inventory demand forecasting, this invention proposes an inventory demand forecasting method with multiple factors. By using the empirical mode decomposition method, the original sequence data is decomposed into new sequences with different characteristics. Different influencing factors and forecasting models are selected according to their respective characteristics for forecasting. This makes up for the excessive reliance of traditional forecasting methods on the sample size of historical data, making the forecast more scientific and accurate while improving the interpretability of the model.

[0032] (2) This invention fully considers the nonlinear and volatile characteristics of agricultural equipment inventory demand, the uncertainty of agricultural equipment from production to inventory, and macroeconomic factors, which is conducive to agricultural equipment enterprises to rationally plan production, reduce inventory waste, and ensure the efficient operation of enterprises. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0034] Figure 2 This is a flowchart of the hyperparameter optimization method of the LSSVM algorithm based on reinforcement learning in this invention. Detailed Implementation

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0036] Example 1

[0037] A method for forecasting agricultural equipment inventory demand influenced by multiple factors includes the following steps:

[0038] Step 1: Obtain historical data on inventory demand and influencing factors from the departmental database, and preprocess the historical data on influencing factors.

[0039] In this step, the influencing factors specifically include: inventory level, order volume, raw material prices, pass rate, productivity, agricultural machinery subsidy policy, farmers' income, fuel prices, arable land policy, and grain prices.

[0040] In this step, the data preprocessing of the historical data of the influencing factors specifically involves: normalizing the historical data of the influencing factors. To avoid the influence of variance and covariance of the units, the normalization process adopts the Z-score standardization method, as detailed below:

[0041]

[0042]

[0043]

[0044] In the formula, y ij y' represents the value of the i-th influencing factor at time j; ij The value is the standardized value; and σ i These represent the mean and standard deviation of the influencing factors, respectively.

[0045] Step 2: Using the Empirical Mode Decomposition (EMD) algorithm, the historical data of inventory demand is adaptively decomposed into multiple intrinsic mode components (IMF) at different scales and a residual component (RES). Specifically:

[0046]

[0047] In the formula, x(t) is the value of inventory demand at time t; imf is the intrinsic modal component, arranged in descending order of frequency; res is the residual component; and j is the number of intrinsic modal components.

[0048] The specific steps of the EMD decomposition method in this step are as follows:

[0049] (1) Based on the upper and lower extreme points of the original signal x(t), draw the upper envelope f using the cubic spline interpolation function. max (t), lower envelope f min (t);

[0050] (2) Find the mean m(t) of the upper envelope:

[0051]

[0052] (3) Let h1(t) be the difference between x(t) and m(t), that is:

[0053] h1(t) = x(t) - m(t)

[0054] If h1(t) satisfies the following conditions: the number of extreme points of h1(t) is equal to the number of zero points or the difference between the two; the mean of the upper envelope of h1(t) is 0, that is, the upper and lower envelopes are symmetrical about the time axis, then h1(t) is the first imf. If these conditions are not met, then process a and b are repeated for h1(t) until the above conditions are met, and it becomes the first imf, denoted as imf1(t).

[0055] (4) Extract the first IMF from the original signal to obtain the remaining part:

[0056] x1(t) = x(t) - imf1(t)

[0057] Let x1(t) be the new x(t), and repeat the above steps to obtain the remaining imf. Stop the loop when the residual function is a monotonic function or is less than a certain threshold. At this point, the original signal is decomposed into a set of eigenmode components and a residual:

[0058]

[0059] Step 3: Use Fourier transform to select the imf with periodic characteristics, input res and the imf with periodic characteristics into the Prophet algorithm for training and modeling, and make predictions on the imf with periodic characteristics based on this prediction model.

[0060] In this step, the Discrete Fourier Transform (DFT) is used to transform the imf to the frequency domain. The resulting modulus exhibits clearly spaced spectral peaks, and the frequency domain plot is discrete, indicating that the imf is periodic. The specific steps are as follows:

[0061]

[0062] In the formula, X(k) represents the data after Fourier transform, and x(n) is the input imf sequence.

[0063] In this step, the Prophet algorithm fits the trend, seasonal, and holiday components in the time series and sums them to obtain the final prediction result. Its basic structure is as follows:

[0064] x(t)=g(t)+s(t)+h(t)+ε t

[0065] In the formula, x(t) is the time series of inventory demand; g(t) is the trend term, representing the non-periodic trend of the time series; s(t) is the seasonal term, generally in annual units; h(t) is the holiday term, indicating whether there is a holiday on that day; ε t The remaining terms.

[0066] In this step, the Prophet algorithm is trained and modeled. The specific steps are as follows:

[0067] (1): Input the timestamps and corresponding values ​​of the time series and train the model;

[0068] (2): Input the length of the time series to be predicted into the trained prediction model;

[0069] (3): Output the future time series trend, including the fitted curve, upper bound and lower bound, etc.

[0070] Step 4: Use Principal Component Analysis (PCA) to select the key influencing factors of the remaining non-periodic IMFs, use the key influencing factors as independent variables, input them into DRL-LSSVM for training and modeling, and use this prediction model to predict the non-periodic IMFs.

[0071] In this step, using influencing factors as independent variables and the remaining IMF data as dependent variables, principal component analysis (PCA) is used to analyze the correlation between influencing factors and IMF data. This reduces the number of influencing factors to a certain extent, removing unnecessary features and generating features with stronger explanatory power as much as possible without losing the vast majority of information. Several principal components are generated, retaining most of the information from the original features, to identify the key influencing factors for each IMF. The specific steps of PCA are as follows:

[0072] (1) Sample matrix acquisition: Let the number of influencing factors be n, and each influencing factor contain q historical data points, then an n×q sample matrix is ​​formed:

[0073]

[0074] In the formula, y ij (i = 1, 2, ..., n, j = 1, 2, ..., q) represents the index of the i-th influencing factor at time j.

[0075] (2) Calculate the correlation coefficient matrix among the influencing factors:

[0076]

[0077] In the formula, r ij (i,j=1,2,...,q) is the data correlation coefficient, r ij =r ji .

[0078] (3) Find matrix R q×q eigenvalues ​​λ i (i = 1, 2, ..., q), and the corresponding eigenvector u i (i = 1, 2, ..., q), the cumulative contribution rate of the first m principal components is:

[0079]

[0080] The top m principal components corresponding to the influencing factors with a cumulative contribution rate exceeding 85% are selected. The component matrices of these m principal components and the original data are substituted into the principal component expression to calculate the score matrix, which is then used as the input matrix of the prediction model.

[0081] In this step, LSSVM is an improved algorithm based on SVM. It introduces equality constraints and a least-squares loss function to transform the optimization problem into solving linear equations, avoiding quadratic programming. While inheriting the advantages of SVM in solving small-sample, nonlinear, and strongly coupled problems, it reduces algorithm complexity and speeds up computation. The accuracy of the LSSVM regression prediction model is closely related to the penalty coefficient and kernel function parameters. If the parameters are too large or too small, overlearning or underlearning may occur, leading to reduced prediction accuracy. This method uses Deep Reinforcement Learning (DRL) to abstract the hyperparameter optimization problem of the LSSVM algorithm into a sequential decision problem, selecting the hyperparameters of the algorithm to be optimized step by step, thus improving search efficiency.

[0082] Deep reinforcement learning algorithms utilize LSTM artificial neural networks to construct an agent. The agent's goal is to find a policy that minimizes the prediction error of the LSSVM algorithm. Its workflow is as follows: In each iteration, the agent selects a set of hyperparameters λ for the LSSVM algorithm with probability P, trains the algorithm on the training dataset, uses the error of the prediction value as the reward value, and updates the policy using the policy gradient algorithm. After multiple training iterations, the agent will select the hyperparameter configuration with smaller error with a higher probability.

[0083] The specific steps for regression prediction based on DRL-LSSVM are as follows:

[0084] (1) In step 4, an m×q principal component score matrix is ​​obtained according to the principal component analysis method. This matrix is ​​used as the input data, and the remaining imf, which does not have periodic characteristics, is used as the output data and input into the LSSVM model.

[0085] (2) The agent in the reinforcement learning algorithm selects a set of hyperparameters λ with probability P as the LSSVM algorithm, trains the algorithm on the training dataset, uses the prediction error as the reward value, and uses the policy gradient algorithm to update the policy. Finally, a set of hyperparameters with the smallest prediction error is selected to obtain the prediction model.

[0086] (3) Input the data to be predicted into the prediction model and output the prediction results.

[0087] Step 5: Add the prediction results from Step 3 and Step 4 together to obtain the final prediction result, specifically:

[0088]

[0089] In the formula, x(t') is the predicted inventory demand at time t'; imf j (t') is the predicted value of the j-th intrinsic mode component at time t'; res(t') is the predicted value of the residual component at time t'.

[0090] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting agricultural equipment inventory demand influenced by multiple factors, characterized in that: Includes the following steps: Step 1: Obtain historical data on inventory demand and influencing factors from the departmental database. Perform data preprocessing on the historical data of influencing factors, which specifically include: inventory level, order volume, raw material price, pass rate, productivity, agricultural machinery subsidy policy, farmers' income, fuel price, arable land policy, and grain price. Step 2: Using the empirical mode decomposition algorithm, the historical data of inventory demand is adaptively decomposed into multiple intrinsic mode components (IMF) of different scales and a residual component (RES). Historical data on inventory demand Perform EMD decomposition, decompose into: In the formula, For inventory demand in The value at time; These are the intrinsic modal components, arranged from highest to lowest frequency. These are the residual components; The number of intrinsic modal components; Step 3: Use Fourier transform to select the imf with periodic characteristics, input res and the imf with periodic characteristics into the Prophet algorithm for training and modeling, and make predictions on the imf with periodic characteristics based on this prediction model. The Prophet algorithm fits trend terms, seasonal terms, and holiday terms in a time series and sums them to obtain the final prediction result. The specific steps for training and modeling the Prophet algorithm are as follows: (1) Input the timestamps and corresponding values ​​of the time series and train the model; (2) Input the length of the time series to be predicted into the trained prediction model; (3) Output the future time series trend, including the fitted curve, upper bound and lower bound; Step 4: Use principal component analysis to select the key influencing factors of the remaining non-periodic IMFs, use the key influencing factors as independent variables, input them into DRL-LSSVM for training and modeling, and use this prediction model to predict the non-periodic IMFs. The specific steps for regression prediction based on DRL-LSSVM are as follows: (1) Obtain an m×q principal component score matrix according to the principal component analysis method, use it as input data, and use the remaining non-periodic imf as output data, and input it into the LSSVM model; (2) The agent in the reinforcement learning algorithm selects a set of hyperparameters with probability P as the LSSVM algorithm, trains the algorithm on the training dataset, uses the prediction error as the reward value, and uses the policy gradient algorithm to update the policy. Finally, a set of hyperparameters with the smallest prediction error is selected to obtain the prediction model. (3) Input the data to be predicted into the prediction model and output the prediction results; Step 5: Add the prediction results from Step 3 and Step 4 together to obtain the final prediction result.

2. The method for predicting agricultural equipment inventory demand based on multiple factors according to claim 1, characterized in that: The specific steps of step 1 for preprocessing historical data of influencing factors are: normalizing the historical data of influencing factors.

3. The method for predicting agricultural equipment inventory demand influenced by multiple factors according to claim 1, characterized in that: In step 3, Fourier transform is used to select the imf with periodic characteristics. Specifically, Fourier transform is used to transform each imf to the frequency domain. The signal with a discrete frequency domain graph is the periodic signal.

4. The method for predicting agricultural equipment inventory demand influenced by multiple factors according to claim 1, characterized in that: In step 4, principal component analysis is used to select the key influencing factors of the remaining non-periodic IMF. Specifically, principal component analysis is used to reduce the number of influences, remove unnecessary features, generate features with stronger explanatory power as much as possible without losing most of the information, and generate several principal components that retain most of the information of the original features, i.e., key influencing factors.

5. The method for predicting agricultural equipment inventory demand influenced by multiple factors according to claim 1, characterized in that: The specific steps in step 5, which involve summing the prediction results, are as follows: In the formula, For inventory demand in The predicted value at any given time; For the first Each intrinsic mode component in The predicted value at any given time; For the residual components in The predicted value at any given time.

Images